CoreWeave said on 30 September that it will add Nvidia’s Vera CPU to its cloud, announcing it at its Fully Connected conference in San Francisco. Nvidia describes Vera as the first CPU built specifically for AI agents rather than for general-purpose computing, and CoreWeave is among the first commercial operators to commit to offering it.
The configuration the two companies are describing is a rack of 128 Vera CPUs holding 11,264 cores, supporting more than 11,000 concurrent agent environments. Each chip carries 88 cores, up to 1.5 TB of LPDDR5X memory at up to 1.2 TB/s, and an NVLink-C2C link to a GPU at up to 1.8 TB/s.
The claimed advantage is in starting, not running
CoreWeave says agent sandboxes start more than three times faster on Vera than on x86, and reports up to 1.8x on code and compile workloads. Both figures come from the vendors and have not been independently benchmarked.
The underlying argument is about what an agent workload actually does. An agent spends much of its time creating a sandbox, running a short task inside it and tearing it down, which stresses process creation, memory bandwidth and the link to the accelerator rather than sustained single-thread throughput. Chen Goldberg, CoreWeave’s EVP of product and engineering, put it as general-purpose infrastructure bottlenecking agentic AI, with Vera the first CPU explicitly designed to accelerate it.
Rubin is the part that is shipping
Separately, a multi-rack Vera Rubin NVL72 cluster — 72 Rubin GPUs and 36 Vera CPUs per rack — entered limited availability on CoreWeave’s cloud on 30 September, with Cognition, the company behind the Devin coding agent, as the first customer running production workloads on it. The standalone Vera CPU service has no launch date and no published pricing.
What to watch
Whether anyone publishes a sandbox-startup measurement on Vera that Nvidia and CoreWeave did not run, and whether the price lands anywhere near x86 instances once it does have one.
